NVIDIA expands DGX Spark to eight-node clusters with 1TB of unified memory for local AI agents
NVIDIA has extended DGX Spark clustering to eight systems, allowing developers to run large language models and distributed AI workloads on locally managed hardware with up to 1 TB of aggregate unified memory. The expansion was announced at GTC Berlin on October 21, 2026, followed by a technical report on October 22 detailing inference performance, networking configurations, model orchestration and distributed training.
The eight-node configuration uses DGX Spark systems equipped with 128 GB of unified memory each. NVIDIA has demonstrated the full 753-billion-parameter GLM 5.3 model running across the cluster, alongside a smaller GLM 5.3 Flash configuration generating approximately 80 tokens per second.
The principal change is not a new processor or an increase in the memory available to an individual machine. It is official support for distributing workloads across eight existing Grace Blackwell systems, using their ConnectX-7 network interfaces to exchange model data.
Eight-node clusters can already be configured manually, but automated eight-node support in NVIDIA Sync Cluster Assistant is scheduled for the end of 2026. The forthcoming NVIDIA Sync Model Launcher is a separate feature intended to simplify model deployment and integration with OpenCode.
NVIDIA has also introduced a 64 GB DGX Spark configuration, available from manufacturing partners starting October 23 at $4,999. That model retains the GB10 Grace Blackwell processor and software stack, but the eight-node, 1 TB demonstration uses the existing 128 GB configuration.
In NVIDIA's Qwen3.8-27B test, eight switch-connected systems delivered 52.58 generated tokens per second, compared with 11.86 tokens per second on one system. A separate small-model training experiment reached 7.79 times the single-node throughput.
··········
WHAT NVIDIA ANNOUNCED AT GTC BERLIN.
The October announcement establishes eight DGX Spark systems as an officially supported cluster configuration for local AI development. Previously, NVIDIA had focused on smaller arrangements, including direct connections between two systems and clusters of up to four nodes.
The larger configuration is intended for developers who need to distribute a model across multiple accelerators, serve several users from shared hardware or run multiple model instances without relying on a remote GPU provider.
Each 128 GB DGX Spark integrates an Arm-based CPU, a Blackwell GPU and unified system memory. The eight systems communicate through ConnectX-7 networking rather than a shared internal memory bus. The cluster therefore does not behave like a single GPU with 1 TB of directly attached memory.
NVIDIA's technical report evaluates serving a large model that cannot fit on fewer nodes, accelerating tensor-parallel inference, routing requests between smaller and larger models, and pre-training a relatively small model across multiple machines.
........
Capability | Eight-node configuration |
Systems | 8 |
Memory per system | 128 GB |
Aggregate memory | 1,024 GB |
Platform | GB10 Grace Blackwell |
Interconnect | ConnectX-7 |
Setup | Manual supported |
Automation | Planned by end 2026 |
........
··········
DGX SPARK HARDWARE AND THE ROLE OF UNIFIED MEMORY.
DGX Spark combines a Blackwell GPU with a 20-core Arm processor comprising ten Cortex-X925 and ten Cortex-A725 cores. The 128 GB configuration uses LPDDR5x unified memory with a 256-bit interface and a published bandwidth of 273 GB/s.
Because the CPU and GPU share a memory architecture, developers can work with models that would exceed the dedicated video memory of many conventional desktop graphics cards. Unified memory does not make memory bandwidth unlimited.
The system includes a ConnectX-7 network interface, 10 GbE Ethernet, Wi-Fi 7 and NVMe storage. NVIDIA documents support for 200 Gbps RoCE connectivity in DGX Spark clustering.
RoCE, or RDMA over Converged Ethernet, enables lower-overhead data exchange for distributed operations. For tensor-parallel inference, nodes exchange intermediate results during execution.
........
Characteristic | DGX Spark 128 GB |
CPU | 20-core Arm |
GPU | Blackwell |
Memory | 128 GB LPDDR5x |
Bandwidth | 273 GB/s |
RoCE networking | Up to 200 Gbps |
Ethernet | 10 GbE |
Wireless | Wi-Fi 7 |
........
··········
HOW MEMORY SCALES FROM ONE TO EIGHT NODES.
One 128 GB system provides 128 GB of local unified memory. Two systems provide 256 GB aggregate, four provide 512 GB and eight provide 1,024 GB.
The memory available for model weights is smaller than the nominal total because runtime processes, activation buffers, communication resources and key-value caches also consume capacity. Quantization reduces weight storage but can affect numerical accuracy and implementation complexity.
A 753-billion-parameter model at four bits per parameter requires approximately 376.5 billion bytes for raw weights before quantization metadata and other allocations. NVIDIA reports that the GLM 5.3 NVFP4 checkpoint occupies approximately 470 GB.
Eight 64 GB systems would provide 512 GB, not the 1 TB demonstrated using eight 128 GB systems.
........
Nodes | Aggregate unified memory |
1 × 128 GB | 128 GB |
2 × 128 GB | 256 GB |
4 × 128 GB | 512 GB |
8 × 128 GB | 1,024 GB |
8 × 64 GB | 512 GB |
........
··········
RUNNING GLM 5.3 WITH 753 BILLION PARAMETERS.
NVIDIA reports running the full 753-billion-parameter GLM 5.3 NVFP4 checkpoint across eight 128 GB DGX Spark systems. Its approximately 470 GB checkpoint is an actual deployment example rather than a theoretical memory calculation.
At concurrency one, NVIDIA reports 31 generated tokens per second. At concurrency four, aggregate decode throughput reaches 51 tokens per second, and at concurrency eight it reaches 67 tokens per second.
The report also gives per-user decode figures of 31, 21 and 18 tokens per second at these concurrency levels. The serving benchmark uses an 8,000-token input sequence per request and MTP3 for generation.
NVIDIA reports approximately 80 generated tokens per second for GLM 5.3 Flash on an eight-node cluster. That is a different model configuration and should not be conflated with full GLM 5.3 performance.
........
Concurrency | Aggregate decode TPS | Per-user TPS |
1 | 31 | 31 |
4 | 51 | 21 |
8 | 67 | 18 |
........
··········
QWEN3.8-27B BENCHMARKS SHOW THE COST OF DISTRIBUTED INFERENCE.
NVIDIA tested a dense Qwen3.8-27B model using NVFP4 precision without speculative decoding or multi-token prediction. A single DGX Spark generated 11.86 tokens per second with 2,635 prefill tokens per second.
Two directly connected systems achieved 21.71 decode tokens per second. Four systems reached 32.82 in a physical ring and 36.34 in the optimized switched configuration.
Eight systems achieved 40.11 decode tokens per second in a physical ring and 52.58 using a switch with RoCEnante. The latter is a 4.43-fold increase over a single node, despite using eight times as many systems.
Prefill increased from 2,635 to 5,105 tokens per second in the best eight-node configuration. The measurements used an 8,192-token input sequence, 256 generated tokens and concurrency one. These are workload-specific figures, not universal model performance.
........
Nodes/topology | Prefill TPS | Decode TPS |
1 system | 2,635 | 11.86 |
2 direct | 3,285 | 21.71 |
4 ring | 4,446 | 32.82 |
4 switch RoCEnante | 4,556 | 36.34 |
8 ring | 4,962 | 40.11 |
8 switch RoCEnante | 5,105 | 52.58 |
........
··········
CONNECTX-7, ROCE AND RING VERSUS SWITCH TOPOLOGIES.
In a physical ring, each node communicates with neighbors. The topology avoids a dedicated switch and can work effectively for collective communication patterns, but direct connectivity between arbitrary nodes is limited.
A switched topology supports more flexible traffic patterns, especially when the cluster is divided between models or independent workloads. It introduces switch hardware cost and configuration requirements.
NVIDIA's eight-node benchmark reached 40.11 decode tokens per second in a ring and 52.58 in a switched configuration using RoCEnante. Because the collective backend also changes, the difference cannot be attributed to the physical switch alone.
··········
ROCENANTE REDUCES COLLECTIVE COMMUNICATION LATENCY.
NVIDIA describes RoCEnante as a Spark-optimized collective backend for tensor-parallel inference. It uses an all-to-all exchange followed by local reduction rather than a conventional ring all-reduce.
This design can reduce sequential communication steps for small messages and modest node counts, at the cost of greater total data movement. It is not a universal replacement for NCCL.
NVIDIA reports eight-node decode scaling of 3.05× without the optimized backend and 4.43× with it in the relevant switched benchmark. That is approximately 45% higher scaling factor under those test conditions, not a guarantee for every workload.
The backend can fall back to NCCL if its startup test fails.
··········
NVIDIA SYNC CLUSTER ASSISTANT: CURRENT AND FUTURE SUPPORT.
NVIDIA Sync provides discovery, configuration and management tools for DGX Spark systems. Cluster Assistant can validate readiness, inspect network connections, plan addresses and configure ConnectX-7 connectivity.
Eight-node clustering is supported with manual configuration. Automated eight-node Cluster Assistant support is planned by the end of 2026. Developers should not assume existing guided workflows already automate every eight-node topology.
Manual setup can involve RoCE settings, addressing, SSH trust, distributed runtime configuration and validation. NVIDIA supplies technical playbooks, but official support does not eliminate this engineering work.
........
Feature | October 2026 status |
2-node clustering | Available |
8-node clustering | Manual configuration |
Automated 8-node Sync | Planned end 2026 |
Sync Model Launcher | Announced for October |
........
··········
NVIDIA SYNC MODEL LAUNCHER AND OPENCODE.
NVIDIA Sync Model Launcher is designed to choose and download a model and serving container appropriate for the detected hardware, then connect the model to OpenCode.
The launcher is also intended to load DGX Spark Skills that provide guidance for troubleshooting, maintenance and documentation. NVIDIA announced availability later in October 2026.
The announcement does not provide independent installation-time measurements, compatibility coverage for every model or a quantified reduction in operational overhead.
··········
MULTI-MODEL ORCHESTRATION WITH NEMO SWITCHYARD.
NVIDIA tested routing between MiniCPM5 2B for classification and simpler requests and Qwen3.8 Flash Next for more demanding work. The eight-node prototype allocated three two-node Qwen replicas and two single-node MiniCPM replicas.
The company reports a score of 91 out of 100 on a 92-case tool-use evaluation, matching its Qwen-only baseline. This limited evaluation does not establish equivalence across all tasks.
When half of requests were routed to Qwen, aggregate throughput was approximately 1.5 times the Qwen-only baseline. When only one-quarter reached Qwen, throughput was approximately three times baseline.
The improvement reflects request routing rather than making the large model itself three times faster. Routing errors can reduce quality when a request classified as simple is actually demanding.
........
Requests to Qwen | Relative throughput |
100% | 1.0× |
50% | Approximately 1.5× |
25% | Approximately 3.0× |
........
··········
TRAINING PERFORMANCE REACHES 97.4% SCALING EFFICIENCY.
NVIDIA evaluated distributed pre-training with nanochat, a model of approximately 896.5 million parameters. Unlike serving a 753B model, this experiment uses data parallelism: each node holds a complete model replica and synchronizes gradients.
A single node processed 8,741 tokens per second. Two reached 17,202, four reached 33,640 and eight reached 68,079. The eight-node result is 7.79 times single-node throughput and 97.4% scaling efficiency.
The test used a global batch of 524,288 tokens per step and 32 sequences per GPU microbatch. This near-linear scaling for a small model should not be extrapolated to frontier-scale pre-training.
........
Nodes | Training TPS | Scaling efficiency |
1 | 8,741 | 100% |
2 | 17,202 | 98.4% |
4 | 33,640 | 96.2% |
8 | 68,079 | 97.4% |
........
··········
THE NEW 64GB DGX SPARK STARTS AT $4,999.
NVIDIA introduced a DGX Spark variant with 64 GB unified memory, retaining the GB10 Grace Blackwell processor, DGX OS and AI software stack. Partner availability began October 23, 2026, with a starting price of $4,999.
Manufacturing partners include Acer, ASUS, Dell, GIGABYTE, HP and MSI. NVIDIA states the configuration can support models up to approximately 100 billion parameters, depending on architecture and deployment.
Two 64 GB systems can be linked over ConnectX-7 for 128 GB aggregate memory. NVIDIA reported up to 1.7 times the performance of one system in a Qwen3.8-27B test.
The $4,999 price applies to the new 64 GB configuration, not an eight-node 128 GB cluster. NVIDIA did not disclose a single complete-cluster price inclusive of switches, cabling and supporting equipment.
........
Product | Memory | Pricing |
DGX Spark 64 GB | 64 GB | Starts at $4,999 |
DGX Spark 128 GB | 128 GB | Varies by vendor |
8-node 128 GB cluster | 1,024 GB | No complete price disclosed |
........
··········
LOCAL AI AGENTS AND THE LIMITS OF ON-PREMISES EXECUTION.
An eight-node cluster can serve as a shared inference backend for multiple developers and agent applications. Larger models can handle complex requests while smaller models handle classification, retrieval and simpler steps.
Local inference can reduce dependence on external inference APIs and provide more control over versions and selected data flows, but it does not automatically make an entire agent workflow private. Connected services and external APIs may still process information.
Key-value caches grow with context length and concurrency, consuming memory otherwise available for model weights. More GPUs do not directly improve a model's factual accuracy or decision reliability.
··········
OPERATING COSTS, NETWORKING REQUIREMENTS AND TRADEOFFS.
Distributed inference requires compatible networking, collective libraries and system configuration. A switched setup may need a dedicated RoCE-capable switch and suitable cables, plus power, cooling and ongoing administration.
NVIDIA does not publish a comprehensive total-cost-of-ownership comparison against cloud inference or dedicated data-center servers. Such an analysis would require purchase price, utilization, electricity, staffing, maintenance and depreciation.
Local ownership may suit sustained predictable workloads or strict operational-control requirements. Cloud capacity may be more economical for short experiments or variable demand.
··········
WHAT NVIDIA'S BENCHMARKS DO NOT PROVE.
The Qwen results use a specific quantization format, sequence length, output length and concurrency. GLM results depend on a particular checkpoint and decoding setup. Switchyard performance depends on routing mix and task difficulty.
The nanochat experiment demonstrates efficient data-parallel training of a sub-billion-parameter model, not equivalent efficiency for frontier-scale training. The report does not establish a standardized full-system cost per token.
These limitations define where the measurements are informative rather than invalidating the reported experiments.
··········
HOW DGX SPARK FITS BETWEEN DESKTOP AND DATA-CENTER AI.
The eight-node configuration offers more aggregate memory and compute than one desktop AI system, while remaining a collection of compact computers that developers can operate locally.
It is not equivalent to a rack-scale data-center accelerator with a purpose-built internal high-bandwidth GPU fabric. ConnectX-7 enables distributed execution, but communication overhead materially limits some inference workloads.
For models fitting on one system, additional nodes may create complexity without proportional benefit. For models requiring several hundred gigabytes of weight storage, clustering can enable deployments otherwise impossible on a single machine.
··········
THE SIGNIFICANCE OF EIGHT-NODE SUPPORT.
Eight 128 GB DGX Spark systems provide approximately 1 TB aggregate memory, enough to host NVIDIA's quantized GLM 5.3 753B checkpoint. The Qwen3.8-27B benchmark demonstrates 4.43× single-node decode throughput in the best tested eight-node configuration.
NeMo Switchyard illustrates throughput gains from routing requests between differently sized models, while nanochat training shows 97.4% scaling efficiency under a separate data-parallel workload.
Manual eight-node clustering is supported now; automated NVIDIA Sync support is scheduled for later. The platform's practical value depends on matching hardware, topology, model and workload rather than assuming that eight nodes deliver eightfold performance.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
